AI Output QA Checklist

AI-assisted work fails in a predictable pattern. The most dangerous failure is the confident wrong fact — an invented stat, a misattributed quote, a feature claim that's a version behind. Style issues are recoverable; a fabricated McKinsey citation in a board paper is not. The skill prioritises factual integrity first, AI language patterns second, tone and structure third — in that order, because that's the order things actually damage trust.

The factual integrity scan flags specific statistics without sources, named studies or reports, historical claims, company-specific numbers, and technical specifics like version numbers or pricing that go stale fast. Each gets a verification tag with the exact claim and why it's suspicious, not a vague "some claims may need checking." Then the AI language pattern pass catches the hedging openers, the exhaustive lists, the false-balance constructions, the hollow transitions, the non-committal conclusions, the delve-and-leverage family, and the sycophantic softeners. Each gets a location and a suggested replacement, not just a complaint.

The tone and structure checks catch the things that make experienced readers feel "AI wrote this" without being able to name why — register shifts within the same piece, perfectly parallel sections, suspiciously comprehensive coverage that hits every angle instead of the angles that matter, missing point of view, the obligatory restating-the-topic intro. The client-fit pass then verifies the work uses the client's terminology, references their specific situation, doesn't contradict last month's deliverables, and is actually actionable for this audience.

The output puts the verdict at the top — PASS, PASS WITH EDITS, or NEEDS REWORK — followed by a flagged table with locations and suggested fixes, plus five quick fixes you can apply in five minutes before sending. Different content types get different standards: a blog post tolerates more AI patterns than a board paper, a transcription needs less style review than a thought leadership piece. The skill asks about audience first because reviewing without that context produces generic advice and misses the actual risks.